Why Choosing a Denoiser Is Harder Than Writing Your Own: Topaz Nyx XL vs NeatVideo vs Open-Source Showdown
Topaz Labs
ABSoft
The author compared four denoising tools for film video: UltraNR (DaVinci Resolve), Topaz Nyx XL, open-source DRUNet/SCUNet (REAL-Video-Enhancer), and NeatVideo. Nyx XL delivered excellent quality but was very slow and slightly sharpened. Open-source models were free but slower and less capable. NeatVideo, after proper tuning, was fast (30-40 fps) and of highest quality, but its interface is complex and bug-prone.
The author, after training his own denoiser 'Archivist' for film grain, tested existing commercial and open-source alternatives on a RTX 4060 Ti system. UltraNR in DaVinci Resolve was simple but ineffective for film. Topaz Video AI's Nyx XL gave near-perfect results with one click, cementing grain while preserving detail, but ran at 1-1.6 fps on 1080p and added unwanted sharpening. The open-source REAL-Video-Enhancer (DRUNet and SCUNet) was free but required Nvidia GPUs and patches; SCUNet was slow (<1 fps without TensorRT) and had less headroom than Nyx XL. NeatVideo, a 20-year-old algorithmic plugin, needed 10-15 minutes of manual tuning but after learning the interface could achieve flawless grain removal at 30-40 fps on the RTX 4060 Ti. However, it was notoriously unstable, frequently crashing during profile setup or rendering. The author concluded that NeatVideo, despite its learning curve, offered the best speed and quality once properly configured.
- Abbreviations
- GPU = Graphics Processing Unit — графический процессор
- CPU = Central Processing Unit — центральный процессор
- CNN = Convolutional Neural Network — свёрточная нейронная сеть
- TensorRT = TensorRT (NVIDIA inference optimization SDK) — TensorRT (оптимизатор вывода от NVIDIA)
- IRL = In Real Life — реальная жизнь (не анимация)
Source: Habr — хаб ML —
original